Agentic AI vs generative AI: the difference is who starts the conversation
· Nativs
Generative AI waits to be asked.
Agentic AI shows up on Monday.
That's the whole difference. Everything else is detail.
The test that separates them
Ask one question: who starts the conversation?
If you have to open a tab, type a prompt and paste in the context — that's generative AI. It's a very clever tool that does nothing until you pick it up.
If it wakes up on its own, goes and finds what it needs, does the job and puts the result on your desk — that's an agent.
A hammer is a tool. A carpenter is not a better hammer.
What generative AI actually does
It generates. That's not a criticism — it's the name.
You give it context, it gives you output. Text, code, an image, a summary. It's extraordinary at that, and most companies have barely scratched what it can do.
But it has no memory of Monday. No access to your ledger. No opinion about whether the number it just produced looks wrong.
It's a brilliant graduate who turns up with no laptop, no logins, and no idea what happened last week. Every single morning.
What an agent adds
Four things, and none of them are the model.
A trigger. Something wakes it up. First Monday of the month. A deal going quiet. An invoice hitting sixty days. Nobody has to remember.
Access. It reads your systems directly — the Xero, the HubSpot, the warehouse. Read-only, but real. It doesn't wait to be handed context in a paste buffer.
Judgement about its own limits. This is the part that matters, and the part most demos skip. A good agent knows when it isn't sure. It does the confident work and escalates the rest.
A record. What it did, when, and why. Because "the AI did it" is not an audit trail.
Take any of those four away and you're back to a chatbot with extra steps.
Why the distinction costs money
Because most "agentic AI" being sold right now is generative AI with a scheduler bolted on.
It runs on a cron job. It has no idea whether its own output is sensible. It never escalates, because it can't tell the difference between a routine month and a month where someone fat-fingered a journal entry.
That thing is worse than no automation. It produces confident work you now have to check line by line — so you've added a review job to replace a doing job.
The question to ask any vendor: what happens when it isn't sure?
If the answer is a shrug, or a confidence score nobody acts on, it's not an agent. If the answer is "it stops and asks you, and here's the log of every time it did" — now you're talking.
The bit nobody says out loud
Agentic AI is mostly not an AI problem.
The model is the easy part. The hard part is the plumbing, the permissions, the rules about when to stop, and someone deciding what "unusual" means for your business. That's why the demos look magic and the rollouts stall.
A generative AI project is a licence.
An agentic AI project is a change to how a job gets done — which is a management problem wearing a technology costume.
So which do you need?
If you want your team to write faster, research quicker and draft better: generative AI. Buy the licences. Teach people properly. That's a real win and it's available this week.
If there's a job that happens every month, follows rules, and eats a day of somebody good: that's an agent. And it will keep paying you back long after the novelty of the first one wears off.
Most companies need both. Very few have thought about which is which.
We build agents that run the job on their own when they're sure, and knock on your door when they're not. Tell us the job — priced before we build it, from £1,000 fixed.